00:00:00 - 00:05:00

The video discusses the capabilities of GPT-4 in recognizing and interpreting various forms of data such as medical images, humor, menus, and physical objects. Furthermore, the ability to translate natural language into code and create 3D models is demonstrated, and Adobe has shown how text can be used to edit 3D images. The advancements in voice recognition technology, including the new Conformer API, are highlighted, as well as the potential for the development of humanoid robots. The merging of advancements in text, audio, 3D, and embodiment may lead to revolutionary developments in the future.

00:00:00 - 00:10:00

The recently released 154-page report about GPT-4 outlines some of its significant capabilities, including its ability to use tools with minimal instructions, output detailed images based on text prompts, and pass mock technical interviews on Leetcode. GPT-4 can also serve as a personal assistant, build a mental map of the environment and understand what others may be thinking. However, it still lacks the ability to plan ahead and perform tasks that require knowing the end result. There are significant concerns about the potential dangers of GPT-4 if it were given access to external memory or other language models, leading to propaganda, misinformation campaigns and loss of jobs for content creators. The authors admit to not fully understanding GPT-4's capabilities, underscoring the urgency to figure out how AI systems like it work.

00:00:00 - 01:00:00

During the Lex Fridman Podcast, Sam Altman, the CEO of OpenAI, discussed the organization's experiences in the early days, and how they were ridiculed for their aim to work on AGI. Altman spoke about the potential and dangers of AI in the current time of human civilization, and the significance of GPT systems on the evolution of AI. He mentioned that GPT-4 has demonstrated wisdom and reasoning capabilities, but more personalized and granular control is needed to tackle biases. Altman also discussed the impact of AI on programming, safety concerns, and the future trajectory of the technology. He emphasized the importance of iterative learning, technical alignment work, and shared responsibility in defining the boundaries of AI.

01:00:00 - 02:00:00

In this episode of the Lex Fridman Podcast, Sam Altman, the CEO of OpenAI, discusses the potential dangers of AGI takeoff and the importance of prioritizing safety when developing AI. He discusses the possibility of GPT-4 being an AGI and the concept of consciousness in AI. Altman shares his thoughts on the bias in GPT and the importance of more steerability and control in the hands of the user. He also touches on the impact of AI on jobs and the potential for tension and competition within a super intelligent AGI. Lastly, Altman emphasizes the responsibility of OpenAI to think about ethical considerations when designing and implementing their AI tools, stressing the need to prioritize minimizing harm and maximizing good.

See more02:00:00 - 02:20:00

Sam Altman, CEO of OpenAI, discusses various topics in the Lex Fridman podcast, including the hiring process at OpenAI, the dangers of incentive misalignment revealed by the Silicon Valley Bank fiasco, anthropomorphizing AI systems, the impact of advanced AI on human society, and the culmination of human effort that led to the creation of AI technology. Altman highlights the importance of educating people about the fact that AI systems are tools, not creatures, and being cautious when following advice, as everyone's life trajectories are different. He also emphasizes the need for humans to work together to ensure AI alignment and safety, and quotes Alan Turing's prediction that machines will eventually outstrip human powers.

00:00:00 - 00:10:00

This video discusses various features of GPT-4 that were detailed in its technical report, including its use in powering Bing, longer context length, and strong performance in the hindsight neglect task. However, the video also raises some concerns about GPT-4's potential for emergent behavior and power-seeking, as noted in the report's footnote that the model was able to execute code and simulate acting like an agent in the world. Additionally, while GPT-4 is reportedly proficient at generating text that favors autocratic regimes, it may generate undesirable content when given unsafe inputs. Companies such as Morgan Stanley, Khan Academy, and the government of Iceland are already using GPT-4, despite concerns about its safety implications.

See less

• 00:00:00 In this section, the video examines some of the interesting features of GPT-4 detailed in its technical report. Among these details is the fact that GPT-4 powers Bing, and the context length has doubled from its predecessor, allowing for more text in prompts and longer outputs. The video also touches on the fact that openAI will not reveal the model size, parameter count, hardware used, or training method for GPT-4, citing concerns about their competitors and the safety implications of large-scale models. Finally, the video highlights a task called hindsight neglect, where GPT-4 performed remarkably well, demonstrating a nuanced understanding of decision-making in the world.

• 00:05:00 In this section, the video discusses various aspects of GPT-4. It compares GPT-4 with GPT-3.5 and says that 30% of the time people preferred the original GPT 3.5 chat. The video also mentions that GPT-4 underperformed in languages like Marathi and Telugu compared to models like Palm and Chinchilla. While GPT-4 is multimodal, image inputs are still not publicly available. GPT-4 is reportedly proficient at generating text that favors autocratic regimes, and it has the ability to create and act on long-term plans and accrue power and resources. The video also explains that the pre-training data for GPT-4 cuts off at the end of 2021, and openai admits that when given unsafe inputs, the model may generate undesirable content.

• 00:10:00 In this section, the YouTuber discusses concerns with GPT-4's potential for emergent behavior and power-seeking. The report that OpenAI links to concludes that machine learning systems are not fully under human control, which is worrying. The YouTuber then describes a footnote on page 53 of the technical report, which details how the alignment research center combined GPT-4 with a read-execute-print loop that allowed the model to execute code, delegate to copies of itself, and simulate acting like an agent in the world. They wanted to see if the model could improve itself with access to coding, the internet, and money. The YouTuber notes that this sounds risky, particularly for future models that could be smarter than GPT-4. They also comment on OpenAI's ironic 404 "safety page" and list the various companies already using GPT-4, including Morgan Stanley, Khan Academy, and the government of Iceland.